TherelativepeakintensityofT2gmode of the
H-NNO device at LRS was 0.77, whereas for
HRS this dropped to 0.68, indicating higher
local proton distribution of H-NNO at HRS
near the Pd electrode. Near-field tip-enhanced
Raman scattering (TERS) was carried on the
H-NNO device at LRS and at HRS near the Pd
electrode (Fig. 2J). Details of control experi-
ments for near-field TERS are provided in fig.
S13. A broad T2gpeak of NNO could be seen
near the Pd electrode at LRS; however, no
such weak peak was detected on NNO near
the Pd electrode for H-NNO at HRS, indicating
relatively higher proton concentration near
the Pd electrode. We used scattering-type scan-
ning near-field optical microscopy (s-SNOM)
at a laser frequency ofw= 952 cm−^1 to image
the local distribution of doping of H-NNO
devices at LRS and HRS. Details of control
experiments for s-SNOM on reference devices
are included in fig. S14. Second harmonic
infrared (IR) (w= 952 cm−^1 ) near-field ampli-
tude images of the H-NNO device at LRS and
HRS near the Pd electrode are shown in Fig. 2,
K and L, insets, respectively. Normalized am-
plitude line profiles of the NNO devices at LRS
and HRS are provided in fig. S15. The first
derivative of the normalized amplitude in-
dicates proton concentration changes near
the boundary between the Pd electrode and
H-NNO channel shown in Fig. 2, K and L. At
HRS, the proton concentration changed over
a longer lateral distance compared with that
of at LRS. The s-SNOM amplitude signal dif-
ferences revealed local chemical composition
differences in H-NNO at different functional
states, which was consistent with Raman re-
sults. Further, the carrier localization length
scale of H-NNO device at HRS was smaller than
that at LRS, as determined from temperature-
dependent electrical transport measurements
(fig. S16). The nanoscale chacterization of de-
vices showed consistent results that the local
proton distribution of H-NNO device at LRS
and HRS near the Pd electrode were differ-
ent. Density functional theory (DFT) calcu-
lations further indicated that differences in
the location of protons could lead to modula-
tion of energy band gap of NNO (figs. S17 to
S20), which is of relevance to different func-
tional states. Nudged elastic band (NEB) cal-
culations showed that the proton migration
barrier could vary from 0.2 to 0.6 eV, depend-
ing on the migration path (supplementary text 1).
Therefore, different local proton distributions
at LRS and at HRS of the H-NNO device could
lead to different functional states.
We also fabricated nickelate devices with
100 nm gap size to demonstrate scalability,
endurance, reproducibility, and ultralow en-
ergy consumption (figs. S21 to S24). In scaled
devices, electrical reconfiguration could berealized with <10-ns electric pulses. The en-
ergy cost for a single synaptic update was
~2 fJ, which is comparable with that in the
brain (1 to ~100 fJ) ( 12 ). To demonstrate com-
patibility with CMOS (complementary metal-
oxide semiconductor) technology, nickelate
devices were fabricated on SiO 2 on Si sub-
strates through both sputtering and ALD
(an industrial technique used to grow high-
quality metal-oxide films for state-of-the-art
electronics), and data are shown in figs. S25
and S26.
To showcase applications of the adaptive
nickelate hardware, we applied the experi-
mental memristive and memcapacitive behav-
iors in RC, a brain-inspired machine-learning
architecture that addresses the issue of train-
ing complexity and parameter explosion, com-
monly observed in traditional recurrent neural
networks (RNNs), by only adapting a simple
output layer. RC explains higher-order cog-
nitive functions and the interaction of short-
term memory with other cognitive processes
( 13 ). Details can be found in supplementary
text 2. To have a baseline comparison, we eval-
uated the performance of our H-NNO device
in comparison with theoretical models ( 14 , 15 )
and experimental reports ( 16 , 17 ) for three
different tasks: MNIST (Modified National
Institute of Standards and Technology database)
digit recognition, isolated spoken digit recognition,SCIENCEscience.org 4 FEBRUARY 2022¥VOL 375 ISSUE 6580 537
Fig. 3. Reservoir computing
simulations with data measured
from nickelate devices.(Ato
D) The simulation results of
reservoirs with H-NNO devices,
compared with theoretical and
experimental memristive models of
reservoirs, demonstrated that a
large and random network of
H-NNO devices could function as a
hardware platform for neuromorphic
computing in solving complex
tasks. The simulation results were
based on the average results of
simulating a sample size of
100 reservoirs with similar hyper-
parameters for each reservoir type
to reduce uncertainty owing to
the stochastic nature of reservoir
networks. As shown in (A) to (C), the
H-NNO reservoirs could achieve
comparable performances on
three tasks with fewer devices. The
performance/device ratios in (D)
indicate that the H-NNO reservoirs,
on average, outperformed the
theoretical and memristive
reservoirs by a factor of 1.4×,
1.2×, and 5.1× for MNIST, isolated
spoken digits, and ECG heartbeats,
respectively.
RESEARCH | RESEARCH ARTICLES